Google Cloud Vertex AI generative inference

2026-10-17 fifth re-read of OpenRouter's Google Vertex (US) row for llama-3.3-70b-instruct: 2.26s P50, 103 tok/s, still $0.72/$0.72. The fourth 2026-10-17 read of the same row was 1.22s / 96 tps, the third was 2.32s / 92 tps, the second was 1.24s / 89 tps, and the earlier one was 1.19s / 92 tps. The second Vertex row at the same price has no latency cell again (the fourth read had 1.40s / 96 tps). measured_* stays the Artificial Analysis long-prompt median (153.6 tok/s, 0.68s first chunk), re-confirmed unchanged this pass from the page summary; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read of OpenRouter's Google Vertex (US) row for llama-3.3-70b-instruct: 1.22s P50, 96 tok/s, still $0.72/$0.72. The third 2026-10-17 read of the same row was 2.32s / 92 tps, the second was 1.24s / 89 tps, and the earlier one was 1.19s / 92 tps. The second Vertex row at the same price now has a latency cell: 1.40s / 96 tps (it was empty on the previous three reads). measured_* stays the Artificial Analysis long-prompt median (153.6 tok/s, 0.68s first chunk), which was not re-fetched this pass; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 third re-read of OpenRouter's Google Vertex (US) row for llama-3.3-70b-instruct: 2.32s P50, 92 tok/s, still $0.72/$0.72. The second 2026-10-17 read of the same row was 1.24s / 89 tps, and the earlier one was 1.19s / 92 tps. A second Vertex row at the same price still has no latency cell. measured_* stays the Artificial Analysis long-prompt median (153.6 tok/s, 0.68s first chunk), re-confirmed unchanged later the same day (summary 153.6 t/s and 0.68s; the comparison table still prints 150 / 0.68); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read of OpenRouter's Google Vertex (US) row for llama-3.3-70b-instruct: 1.24s P50, 89 tok/s, still $0.72/$0.72. The earlier 2026-10-17 read of the same row was 1.19s / 92 tps. A second Vertex row at the same price still has no latency cell. measured_* stays the Artificial Analysis long-prompt median (153.6 tok/s, 0.68s first chunk, re-confirmed unchanged later the same day from the embedded byPromptType.long JSON at 153.59 / 0.678); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 Artificial Analysis re-read. The provider page's embedded per-provider JSON (byPromptType.long, the long-prompt median) now gives Google Vertex 153.59 output tok/s at 0.678s median first chunk, up from the 149.9 tok/s recorded on 2026-10-16. TTFT is unchanged at 0.68s, still the lowest of the providers on that page, so measured_ttft_ms stays 680. The page disagrees with itself: its summary prose says 153.6 while its comparison table still prints 150, so the scalar takes the embedded JSON rounded to one decimal and the old 149.9 stays in this note rather than being deleted. Price is untouched at $0.72/$0.72 from Google's own pricing page; the same JSON restates it. OpenRouter's Vertex row the same day (1.19s P50, 92 tok/s) remains a different workload and is not promoted. 2026-10-17: OpenRouter's Google Vertex row for llama-3.3-70b-instruct now publishes a latency cell on the US-flagged row — 1.19s P50, 92 tok/s, still $0.72/$0.72 (a second Vertex row at the same price has no cell). The 2026-10-16 read had no latency cell. measured_* stays the 2026-10-16 Artificial Analysis median (149.9 tok/s, 0.68s first chunk, the lowest TTFT of 11 providers on that page). The OpenRouter P50 of 1.19s / 92 tok/s is a different workload and is not promoted over it; the older OpenRouter 2026-10-14 P50 was 4.44s / 98 tok/s. Price fields stay $0.72/$0.72 from Google's own pricing page. 2026-10-16 refresh. The $0.72/$0.72 price is now sourced from Google's own pricing page, not only from OpenRouter's pass-through row. Measured speed moved: OpenRouter's 2026-10-14 P50 for the Vertex row on llama-3.3-70b was 4.44s latency / 98 tok/s; Artificial Analysis on 2026-10-16 measures Vertex at 0.68s median first chunk (the LOWEST TTFT of 11 providers) and 149.9 output tok/s. Both are kept: the old pair is OpenRouter production traffic, the new pair is AA's benchmark, and the record now carries the third-party measurement because measurement_source is a single enum. Previous measured values: ttft 4440 ms, tps 98. Two services under one endpoint family: first-party Gemini (per-model global/non-global tables, non-global +10%) and the partner-model MaaS catalog where open weights (Llama, DeepSeek, Qwen, GLM, gpt-oss, Mistral) are served on Google's own infrastructure with Google-set prices. Measured figures are OpenRouter's 2026-10-14 P50 on Google Vertex's llama-3.3-70b endpoint: 98 tok/s but 4.44s latency - the slowest TTFT on that table, an order of magnitude behind Groq, which is the finding that matters for latency-sensitive callers. Vertex also sells tuned-model endpoints at 1.5x base price.